firehose> #llmops

Incremental vs Discontinuous Tasks

A taxonomy of intellectual tasks the 2023 Contrary Research report borrows from Microsoft Research's "Sparks of AGI" paper: incremental tasks are solved sequentially or gradually, with progress made one word or sentence at a time (summaries, factual Q&A, composing within a known form) — a shape that suits autoregressive next-token prediction. Discontinuous tasks require a creative leap or a new framing of the problem (jokes, novel scientific hypotheses, new styles) — a shape the report states 2023-era models were bad at. The taxonomy's use is routing: it predicts which delegations to a model will disappoint before you try them, and it explains the layman's whiplash ("computers aren't so smart after all") as a category error — asking a discontinuous question of an incremental engine.

The report's second-brain application: making incremental connections across a large store of existing notes is squarely in the incremental class, so an LLM over your knowledge base plays to the model's strength even when open-ended creativity does not.

Claims


Linked from